Nvidia is now AI's central bank. The double exposure is the risk.

Nvidia is no longer just selling the picks and shovels of the AI boom. It is now backstopping the mines themselves.

Three deals this summer have turned the chipmaker into a private lender of last resort for the very projects that buy its chips. In July, Nvidia promised to top up customers' income if their data centres miss set revenue targets, an earnout-style guarantee rarely seen in chip sales. Earlier in August, it revealed a plan with six Wall Street firms to mobilize more than $500bn of AI infrastructure investment, by guaranteeing the value of the equipment Nvidia itself sells into those projects and underwriting as much as a quarter of some investments' costs. Mid-August brought a third: an agreement to backstop up to $105bn for a single vast Ohio facility that will consume its chips.

A central bank, in plain terms, is the institution a market turns to when private lenders will not lend: it prints money against its own balance sheet to keep credit flowing. Nvidia is now doing the same thing for AI infrastructure, only with chip-purchase commitments, revenue guarantees, and equipment-value insurance instead of interest rates. The mechanism matters because, in a real downturn, Nvidia would be exposed twice: once as a chip seller whose volumes and prices both fall, and again as a financier of the projects whose revenue shortfalls it has promised to absorb.

In the late 1990s, Cisco and Lucent lent billions to telecoms providers that then collapsed when demand fell short. The line between "enabling demand" and "creating it" is a thin one, and the moment a chip vendor starts underwriting its customers' revenue, the company has stepped across it. Jay Goldberg of Seaport Research Partners put it that way to the Economist: Nvidia is walking a fine line between the two, "getting pretty close" to crossing it but not over.

The newer fact is the scale. Nvidia hit a $1trn valuation after 30 years, $2trn nine months later, and $5trn less than two years after that; it is now worth around $5.4trn and is the world's most valuable company. Some analysts cited in the briefing expect roughly $1trn in annual revenue by 2029, with next-year sales expected almost to double. Those are analyst projections, not reported figures, and the gap between projection and outcome is exactly where the new financing structures sit. Each structure lowers the cost of a build for Nvidia's customers by making the chip vendor the party absorbing some of the demand risk; for Nvidia, the equipment is the same chips it would have sold anyway, plus a contingent liability sitting next to them on the balance sheet.

Michael Burry, the investor who made his name betting against the mortgage-backed securities at the heart of the 2007–09 financial crisis, has gone public with the obvious follow-up question: what happens if AI chip demand grows more slowly than expected, if supply increases, and if prices fall? In that case, the data centres Nvidia has helped finance would be the first to pull back on orders, and Nvidia would still owe the revenue top-ups and equipment-value guarantees it has already signed. The order book for chips and the credit book for the same chips would be working in opposite directions.

Nvidia has chosen to be the credit backstop of an infrastructure cycle whose demand is still being built. If the cycle holds, Nvidia earns financing income on top of chip sales. If it does not, the same balance sheet that funded the boom is on the hook for the bust, and the chip vendor has become the bank. The next test arrives whenever one of these top-up or guarantee structures is first called on, and the reporting on whether Nvidia's own quarterly filings treat those commitments as off-balance-sheet risk or as ordinary revenue support will tell the market whether the new role is being priced at all.